* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download An Empirical Study of Taiwan Bond Market Based on Nonlinear
Pensions crisis wikipedia , lookup
Business valuation wikipedia , lookup
Internal rate of return wikipedia , lookup
Financial economics wikipedia , lookup
Financialization wikipedia , lookup
Credit rationing wikipedia , lookup
Interbank lending market wikipedia , lookup
Adjustable-rate mortgage wikipedia , lookup
Continuous-repayment mortgage wikipedia , lookup
History of pawnbroking wikipedia , lookup
Yield curve wikipedia , lookup
An Empirical Study of Taiwan Bond Market Based on Nonlinear Dynamic Model 1. Corresponding Author: Tsangyao Chang, Ph.D. Professor of Economics and Finance, Chairman & Director, Department of Finance, Feng Chia University, Taichung, Taiwan. TEL: 886-4-2451-7250 ext. 4150. FAX: 886-4-2451-3796. E-Mail: [email protected]. 2. Shu-Chen Kang, Ph.D. candidate. College of Business, Feng Chia University, Taichung, Taiwan, TEL: 886-4-2451-7250. ext. 4217 E-Mail: [email protected] FAX: 886-4-2451-6885. An Empirical Study of Taiwan Bond Market Based on Nonlinear Dynamic Model Abstract This paper examines long-run dynamic adjustment of the term structure of interest rate using Taiwan’s government bond interest with different maturities for the period of 2000 to 2003. We employ a methodology that permits threshold and the momentum-threshold adjustment to test asymmetry unit-roots and cointegration. Specifically, we examine if the term structure of interest rate is consistent with expectation theory by using non-linear methodology. To compare with previous research, we assume that the dynamic adjustment of yield spreads in different maturity bonds. The result supports the expectation theory of the term structure of interest rate with dynamic adjustment. It could be resulted in bias by using symmetry adjustment assumption to build the term structure of interest rates. Furthermore, whenever interest rates increase or decrease, we find the effect of asymmetric price transmissions between different maturity bonds both in the short and long run. But the result is not significant when interest rates increase. We use asymmetry error-correction model to catch up the dynamic adjustment of interest rates. Keyword: Term Structure of Interest Rates, Threshold Autoregressive Model (TAR), Momentum- Threshold Autoregressive Model (M-TAR) -1- 1. Introduction Since new trading systems and new financial products were introduced to Taiwan bond market in 2000, bond trade volume witnessed a sharp increase. The daily trading volume has far surpassed the turnover of trading in Taiwan stock market. However, the fact that Taiwan bond market is just confined to institutions such as banks, securities and insurance company often leaves the bond market ignored by the public. Since interest rate constitutes one of major factors that influence the prices of the financial instruments in the financial market, the fluctuation in interest rate in the bond market is regarded as a leading indicator for the trend of interest rate. Therefore, for the individual, enterprise or financial institution, a good command of the long-term and short-term change in interest rate can contribute to reducing business risks. Term structure of interest rate indicates the relations of yield rates of the bonds with different durations to their maturities. In accordance with term structure of interest rate, the theoretical price of a bond can be judged at any place and time to avoid the risk in investment portfolio and evaluate investment performance. In addition, the term structure of interest rate reflects all market participants’ expectations on the interest rate and inflation in the future. As far as the policy maker is concerned, term structure of interest rate can serve as a tool for analyzing monetary policy. Fisher (1930) firstly proposed expectation theory. According to the theory, investors’ expectations on the future spot interest rate will influence the current long-term interest rate. The theory was further developed by Lutz (1940) who believed that the relations between the yields with different maturities were subject to the investor expectation on the future interest rate. However, expectation theory has been playing an important role in the empirical study on the term structure of interest rate. The theory held that the current long-term interest rate is equivalent to the expected short-term interest rate in the future and the premium that reflects the liquidity and preference at the maturity. In conducting variance bound test, Shiller (1979) discovered that it wasn’t consistent with the hypothesis that long-term rate of interest was the mean value of the expected short-term interest rate and the fluctuation is little and expectation theory is not applicable when the yield rate of the long-term bond fluctuated more violently than the interest rate of the short-term bond. Campbell and Shiller (1987) argued that the necessary condition for the term structure of interest rate to fit in with the expectation theory is the existence of the co-integration -2- between the long-term and short-term rates of interest. That is to say, the premium on the yield rates in each period is not featured by unit-root under the circumstances of long-term balance. In most studies on the unit-root test and co-integration, it is hypothesized as a linear adjustment, as in the case of Mankiw and Miron (1986), Campbell and Shiller (1987), Hardouvelis (1988). Their empirical results failed to support the expectation theory. They held that the negligence of the time-vary premium in the regression formula accounted for the failure to forecast the future interest rate through the spread. Mankiw and Miron (1986) and Hardouvelis (1988) held that the change of the expected short-term interest rate in the future could be forecasted by means of the spread as a result of the structural shift in the monetary policy. But the long-term interest rate defies easy prediction. The shift in the policy structure may be related to the time-vary premium. Gerlach and Smets (1997) conducted a study on the behaviors of long-term and short-term interest rates in 17 countries and discovered the possibility of predicting the short-term interest rate by way of the spread and the presence of time-vary premium,which confirmed the falsehood in the hypothesis of predicting the term premium through expectation theory. But the time-vary premium may explain the expectation theory which is usually unaccepted in the American literature. Iyer (1997) held that the general reason for rejecting expectation theory might lie in the term premium’s variation with the time as well as the error of the market participants. As was shown in the previous literature, the trend of interest rate would vary with the economic climate. During two oil crises and the period from 1978 to 1982 which witnessed a shift in the monetary policy, the economy remained terribly unstable and the interest rate rose and fluctuated violently. In addition, inflation and business cycle both generate the variation in the interest rate. The trend of interest rate is invulnerable to the economic climate, so the hypothesis of unchanged premium developed in the empirical study within the framework of the traditionally practiced expectation theory cannot present a panoramic view of the shift in interest rate. The term premium of the interest rate is subject to the total economic conditions, which will results in the positive and negative dynamic adjustment after the deviation from the long-term balanced rate of interest. Nevertheless, a growing number of studies revealed the non-linear asymmetrical adjustment of time sequence of many global variables, as evidenced by the industrial production index, effective exchange rate and unemployment rate (Neftci, 1984; Delong and Summers, 1986; Falk, 1986; Sichel, -3- 1989, 1993; Terasvirta and Anderson, 1992; Beudry and Koop, 1993). Also, the non-linear adjustment was found in the financial literature (Kragler and Krugler, 1993; Obsfeld and Taylor, 1997, Coakley and Fuertes, 2001). Balke and Fomby (1997) found that the long-term and short-term interest rates in the U.S. are subject to non-linear asymmetrical adjustment. They further pointed out that when variables present asymmetrical adjustment, the traditional linear co-integration model will generate insufficient test power and estimation error. In contrast to the traditional hypothesis that variables were subject to linear adjustment (Engle and Granger (1987), Johansen (1996), Tong (1983) employed Threshold Autoregressive Model (TAR) to explore the asymmetry of the variables and Enders and Granger (1998), Caner and Hansen (2001) further utilized the Momentum-Threshold Autoregressive Model (M-TAR) to explain the variable’s asymmetric adjustment featured by increment or decrement. Enders and Granger (1998) spotted the insufficiency in test power in the process of traditional linear unit-root test and co-integration test when asymmetric adjustment happens to the economic variables. In the related studies on term structure of interest rate, Rudebush (1995) observed the asymmetry in the probability between increment and decrement in the yield curve. In applying M-TAR, Enders and Granger (1998), Enders and Siklos (2001) discovered that when the fluctuation in long-term interest rate surpassed that in short-term interest rate, the interest rate will be restored at the rate of asymmetric adjustment to the original equilibrium, namely, the non-linear long-term co-integration. Van Dijk and Franses (2000) found the same was true of the interest rate in Holland. In addition, Tzavalis (1999), Andreou et al. (2000), Coakley and Fuertes (2002) discovered the asymmetrical adjustment to the threshold value in British interest rate. Since the year 2000, the global political instability and frequent terrorist attacks led world economy to a sharp recession. In addition to the September 11th terrorist attack in the U.S. and Hurricane Danny, the economic slump and continuously weakened external demand in Taiwan led to the fallen domestic consumption and investment. For arresting this trend, the central bank responded actively to stimulate economic recovery and expand exports. The discount rate was lowered 13 times during the period from the end of 2000 to 2002. At the same time, interest rates in countries such as U.S. and Japan was lowered more than 10 times. Throughout the world, the decreasing width of the interest rate chalked the lowest record while the -4- interest rate was at the unprecedented low level. If the linear symmetry was hypothesized in the adjustment process, the error would happen to the model prediction. In view of this, one of the purposes in the current study is to adopt the adjustment to the premium of interest rate and explain by means of dynamic asymmetrical model the expectation theory on the term structure of interest rate in Taiwan bond market. It is worth noting that the previous Taiwanese researches on term structure of interest rate concentrated on the empirical study on the interest rate in the secondary market of commercial paper, as in the case of Shen(1993), Chuang and Duan(1996), Lin, Hong, and Guan (1998). The commercial paper (CP) is a short-term money market tool, with the term of 360 days at its maximum. Therefore the interest rate of CP as a research target is quite suitable for the research on the short term structure of interest rate. In the previous CP studies, the research target was the short-term part of the interest rate curve. The present study will analyze Taiwan bond market to supplement the research on the long-term interest rate curve. The second purpose of the current study is to explore term structure of long-term interest rate in Taiwan bond market by considering its unique trading characteristic. Enders and Granger (1998), Kuo and Enders (2004) respectively discovered the non-linear dynamic adjustment to the premium of the interest rate in the U.S. and Japan. The third purpose is to adopt TAR to explore whether the premium of the yield rate will show asymmetrical adjustment in the long-term balanced Taiwan bond market. Except for the preface, the study falls into five sections. The second section will provide a brief discussion on the expectation theory of the term structure of interest rate, the third section will estimate the non-linear term structure of interest rate through asymmetrical test, the fourth section will brief the possible reasons for asymmetrical adjustment of the term structure of interest rate in Taiwan, the fifth section will provide the description of data characteristic and empirical result,and the sixth section is the conclusion. 2. Expectation theory and cointegration In the rational expectation hypothesis, if an investor is risk-neutral, then expected excess return of the bond is equal to instantaneous interest rate; if an investor is risk averse, then he will obtain the return on the premium, in addition to the return on the instantaneous interest rate. Hall et al. (1992) expressed the expectation theory as follows: -5- 1 n R(n, t ) Et R(t i 1) L(n, t ) n i 1 (1) Where, R(n, t ) refers to the premium of the bond in a period, E t rational expectations based on the obtainable information at the point of t, L(n, t ) term premium which indicates the obtainable risk premium of long-term bond at the point of t. In the risk-neutral hypothesis, L(n, t ) is zero. Campbell and Shiller (1987) held that the expectation theory on term structure of interest rate is valid, then spread can be used for predicting interest rate, which is equivalent to saying that the term premium is fixed (or stationary). Suppose the yield rate of bond shows I(1) status in time sequence, which represents that the yield sequence will be stationary after the first order difference. The different yields to maturities may exist in co-integration and Equation (1) can be re-expressed as follows: S (n, t ) R(n, t ) R(1, t ) 1 n 1 i j Et R(1, t i) L(n, t ) n i 1 j 1 (2) Where, S (n, t ) refers to spread while R(n, t ) R(n, t ) R(n, t 1) If R(1, t ) and term premium L(n, t ) are stationary, then spread S (n, t ) R(n, t ) R(1, t ) will also be stationary. Under such circumstances, the expectation theory can be applied to explain the co-integration between different term yield rates. Campbell and Shiller (1987) and Shenzhonghua(1993) respectively utilized dual variable vector autoregressive model to study the combinations of both long-term and short-term spreads, for the purpose of exploring the presence of co-integration in the long-term and short-term interest rates in both the U.S. and Taiwan. The research result finally disaffirmed the expectation theory and presented the possible reason for the invalidity of the theory----the existence of noise irrelevant to the sequence or seasonal interferences. Therefore, the expectation theory cannot be applicable until the co-integration between different paired yields to maturity. 3. Testing for Asymmetry Adjustment 3.1 Threshold unit-root tests The unit-root test model for testing term structure of interest rate which was -6- developed by Dickey and Fuller (1979) is shown below: S t 0 1 S t 1 t (3) Where,Δ refers to differential operation factor, S t the spread between any two interest rates, t residual error and t ~ N (0, 2 ) white noise. If S t sequence is in stationary state, then the probability of the random process of its time sequence won’t vary with the passage of time and any impact in either the positive or negative direction will exert temporary effect and will gradually diminish with the time to return to the previous mean value, without producing any long-standing effect. In addition to the DF test, unit-root test also has ADF (Said and Dickey,1980), PP (Phillips-Perron,1988), and KPSS (Kwiatkowski, Phillips, Schmidt and Shin, 1992). However, the abovementioned methods for unit-root test are based on the hypothesis that the time sequence is subject to linear adjustment. Pippenger and Goering (1993), Balke and Fomby (1997), Enders and Granger (1998), Enders and Siklos (2001) proposed the idea that if asymmetry existed in a sequence, then traditional unit-root test and co-integration test would generate the lower power of the test result. Hence, when term structure of interest rate is asymmetrical adjustment, the deviation may result from the application of the traditional unit-root test. Enders and Granger (1998) took Threshold Autoregressive (TAR) to test the non-linear asymmetrical adjustment. The test method is listed as follows: S t I t 1 (S t 1 ) (1 I t ) 2 ( S t 1 ) t (4) Where, I t is Heaviside indicator function, which can be expressed as follows: It 1 0 if St 1 if St 1 (5) When time sequence is adjusted to be symmetrical, then we cannot disaffirm null hypothesis 1 2 and when threshold value is zero Equation (3) is a special case of Equation (4). On the other hand, the sufficient condition for the spread S t to be stationary is 2 ( 1 , 2 ) 0 . When the variant of residual error t is great in quantity, only one i may exist between -2 to 0 while the other j will be zero. Even within the unconvergible unit-root area ( j =0) , t can just be transferred to -7- the converged area. The non-linear null hypothesis is that test F has a unit root, namely, 1 2 0 . The maximal test value t of some ρ is referred to as t-Max, while the minimal test value is referred to as t-Min and F test value . Enders and Siklos (2001) proposed the necessary condition for nonlinear regressive convergence: 1 and 2 shall be in the negative. Under such condition, t-Max can be directly used for verifying the existence of unit root. However, when either 1 or 2 is not negative, shall be used to give a statistics of the test value. Test value will be of more power than of t-Max and t-Min. Therefore, if it is hard to refute the hypothesis in which the sequence represents Random Walk, and ω in Equation (4) is the threshold factor of the long-term balanced value of spread, namely, when S t 1 is above the threshold value in the adjustment process, then the term to be adjusted is 1 (S t 1 ) . When S t 1 is below the threshold value , the term to adjusted is 2 (S t 1 ) . Additionally, when the spread S t is in stationary state, Tong (1983) proposed that when null hypothesis of 1 2 0 is refuted, test value F can be reused to test the presence of symmetry ( 1 2 ) in the adjustment process. However, we have no accurate idea of the special property of nonlinear relation of the sequence. Enders and Granger (1998) and Caner and Hansen (2001) developed Momentum-Threshold Autoregressive Model (M-TAR). In the process of sequence adjustment (possibly dynamic), S t 1 is subject to variation. Equation (5) will be therefore expressed this way: It 1 0 if S t 1 if S t 1 (6) Enders and Granger (1998), Enders and Siklos (2001) pointed out that when sequence adjustment was asymmetrical, if the adjustment is continuous, as compared with the adjustment in the other direction, and momentum will be generated in the specific direction, this model is referred to as M-TAR. Generally speaking, when positive deviation continuously surpasses negative deviation, TAR model is capable of catching the deep process. With regard to M-TAR model, it is the adjustment with S t 1 taken as the threshold value for auto-regression, which determines that it can catch the sharp process. M-TAR model can be employed to judge the existence of unit root in the same way TAR model is applied. The statistic value of F * is used to verify null hypothesis 1 2 0 . If the data characteristic is in asymmetrical adjustment, we won’t predetermine to apply TAR or MTAR. Instead, either AIC or -8- SBC judgment is employed to choose the most appropriate model. If the self relation of residual error is considered in Equation (4), then Equation (4) can be revised as follows: q S t I t 1 (S t 1 ) (1 I t ) 2 ( S t 1 ) S i t i t (7) i 1 Prior to the verification, Equation (7) will be first regressed to figure out the residual error term Ŝ t of the spread S t . The explainary variable in the regression equation is ( S t 1 ) . The following regression equation is worked out in accordance with equation (5) or Equation (6). Sˆ t I t 1 Sˆ t 1 (1 I t ) 2 Sˆ t 1 q Sˆ i t i i (8) i 1 Next, statistic value F is employed to verify null hypothesis 1 2 0 . If the refusal to null hypothesis represents the absence of the sequence, then critical value verified by the statistics of F can be referred in Table 1 developed by Enders and Granger (1998). The verification on the symmetrical adjustment ( 1 2 ) depends on the general F distribution. Enders and Siklos (2001) pointed out M-TAR model is quite good at catching the phenomenon when the time sequence is in sliding change while Caner and Hansen (2001) also pointed out that when the time sequence is non-stationary, M-TAR model is better at catching asymmetrical dynamic adjustment. 3.2 Threshold Cointegration When interest rate has unit-root characteristic or the spread S t is in stationary state, cointegration model can be utilized to check whether the trend of interest rate will exist long. The term structure of the long-term interest rate is expressed as follows: R(n, t ) 0 1 R(1, t ) t (9) When the residual error t in Equation (9) is stationary, it represents the co-integration relation between different interest rates. Enders and Siklos (2001) rewrote Equation (9) in the same form of Equation (8). q ˆ t I t 1 ˆ t 1 (1 I t ) 2 ˆ t 1 ˆ i i 1 Where, -9- t i t (10) It 1 0 if ˆ t 1 if ˆ t 1 or It 1 0 if ˆ t 1 if ˆ t 1 In the cointegration equation, we still make use of test values and * of the statistics F in both TAR and M-TAR models to judge the existence of co-integration between different interest rates. The inter-variable adjustment is symmetrical, namely, null hypothesis 1 2 cannot be refuted, then Engle-Granger cointegration equation is a special case of Equation (10). Enders and Siklos (2001) and Enders and Dibooglu (2002) pointed out the power of the asymmetrical adjustment of * exceeds that of Engle-Granger test. Meanwhile, when there is an increase in asymmetrical effect, test value is better than Engle-Granger test in terms of power. Generally, threshold value is unknown. Chan’s (1993) method can be applied to find out the minimal value of the square term of residual error and figure out the value of threshold value 1 . Enders and Siklos (2001) suggested the application of Akaike information criterion (AIC) or Bayesian information criterion (SBC) to select TAR or M-TAR model. 4. Reasons for asymmetrical adjustment in Taiwan The central bank concerns the stability in commodity prices while formulating monetary policy. Stlvester and De Haan (1996) insisted that the monetary policy can serve multiple goals, with the stability in commodity prices as the principal responsibility. On the other hand, the central bank shall value the adjustment to the monetary policy for inflation more than that for deflation. The same is true of central bank’s attitude towards exchange rate. It shall tolerate currency appreciation rather than depreciation, especially when economy cycle witnesses depression and slump. For the sake of stimulating the stagnant economy and expand aggregate demands, central bank tends to create favorable monetary atmosphere to lower interest rate and boost economy. But the monetary policy won’t bring the benefits immediately after implemented, there is always a lag effect. So central bank usually adopts more conspicuously effective measures to offset the negative influences resulting from the economic slowdown. Under the circumstances of long-standing balance, varied shifts in monetary policy possibly occur due to the central bank’s expectations on the Enders and Siklos(2001) and Chan’s (1993)methods are to arrange residual error term in the order from small to big. The first and last 15% will be removed and the middle 70% of residual error term selected. The sum of squares of the residual error is minimized before the value is figured out. 1 - 10 - recession and inflation. Since the second half of the year 2000, bubble economy plagued the whole world including Taiwan, whose economy fell into recession and total economic data manifested a poor performance. Additionally, there was a pessimistic expectation among governmental authority and market participants that the economic degeneration would still exist. These factors combined to drive the central bank to keep lowering discount rate as U.S. Federal Reserve Bank did. Regardless of the somewhat exaggerated seriousness in inflation, as evidenced by the deflation, the central bank still pursued the expansive monetary policy (for instance, lower the discount rate and deposit reserve ratio.) for the purpose of preventing the economy from worsening and exercising the currency conduct mechanism through financial institutions to divert the funds to enterprises and the public. These efforts were made in expectation of encouraging the investment of private funds to check the falling interest rate. Simultaneously, the capital cost was reduced to attract manufacturers to industrial investment through banking system in order to effectively stimulate investment and consumption. Therefore, the policy implementation may constitute a factor for nonlinear dynamic adjustment. During the current research period, central bank showed a good flexibility in adjusting its monetary policy, as is evidenced by the sharp decrease in the interest rate within only 1 year. The adjustment to term structure of interest rate is possible to produce asymmetrical adjustment. 5. Empirical Results Enders and Granger (1998), Enders and Siklos (2001) discovered that there was an asymmetry in the adjustment to interest rate in the U.S. In consideration of the fact that Taiwan often follows the U.S monetary policy, it is thus safe to assume that the adjustment to term structure of interest rate in Taiwan also appears asymmetrical. The sample data selected for the current study is the information on the yield rates of government bonds with the terms of 10 years, 15 years and 20 years in Taiwan government bond market. The research lasted 4 years from 2000 to 2003. The information sources are Reuter and OTC (R.O.C). Taiwan government bonds were still in the bud at the early period of 1990s and the market was inactive at the early stage of development and faced with many restraints such as insufficient distribution, unstable supply, few market participants, lack of market maker and trading activities featured by buy-and-hold. Until 1999 20-year government bond was initiated. Nevertheless, thanks to the installation of electronic transaction platform for - 11 - government bonds at OTC in 2000, the exclusive computer network contributed to the rationality and transparency in price formation for the bond market as a whole. So the trading volume in Taiwan bond market began to grow rapidly since the year 2000 and the daily volume in the bond market has surpassed stock market by a comfortable margin. The representative terms for the bonds in the bond market are 10 years, 15 years and 20 years. The transaction over these three bonds accounts for more than 90% of the volume of trading total government bonds. In addition, the currently distributed government bonds are featured by these three bonds. Therefore, the present study chose these three bonds as the target for analyzing the adjustment to the term structure of interest rate. It is worth noting that despite numerous issues, Taiwan government bond market doesn’t draw on the re-opening mechanism, which leads the recently issued bonds to be favored in the market while the government bonds due on maturity grow unpopular at the same period. The rapid replacement of leading government bonds makes the market participants develop the inclination to discard the old for the new. Therefore, in selecting the government bond yield rate, the old bond is continuously replaced2, by the new bond with the same term so that the bond yield rate is the latest turnover yield in the market. In this way, the price distortion3.resulting from the issue of new bonds can be avoided. Figures 1 and 2 provide the trend of three kinds of government bond yields and the paired spread trend. It can be seen from Figure 1 that interest rate shows a tendency of sharp slowdown since the year 2000. Until 2003 the government bond yield began to increase with the gradual rise in the economy and the market expectation on the possibly raised interest rate by central bank. On the other hand, in Figure 2---the paired spread trend shows that the sharp rise in the spread seems more continuous than the sharp decrease in the spread. It can be observed from the Figure that it is feasible to adopt TAR or M-TAR model to estimate the spread adjustment. The following is the analysis of the empirical result. 5.1 Unit-root test on asymmetry 2 Since the mainstream bonds are active in transaction and dealer quotation is active, the trading volume accounts for lion’s share of the purchase and sell. Quite opposite, other bonds with different durations is little traded and the trading volume is comparatively meager. 3 The traditional study on term structure of interest rate is based on the analysis of the interest rate of zero-coupon. However, the method is taken to choose the research target due to the imperfect Taiwan bond marketsystem . - 12 - First of all, Table 1 presents the basic statistical information on the interest rates of the 10-year, 15-year and 20-year government bonds. It can be discovered in the table that Jerque-Bera test shows the abnormal distribution of the interest rates of these three bonds. In addition, the unit-root test on ADF, PP and KPSS revealed the unit-root characteristic of these bonds, that is to say, the sequence is featured by random walk. When the yield rate is stationary during some terms, it is a necessity to judge whether the spread is in stationary state, namely, cointegration exists in the interest rate. <Insert Table 1> To check whether the spread is of unit-root nature, the current study adopts the traditional linear Augmented Dickey-Fuller test and asymmetrical TAR and M-TAR unit-root test. Table 2 provides the test results. The first and the second column are the TAR and M-TAR test results when the threshold value is zero while the third column is the result of Augmented Dickey-Fuller test. Additionally, the threshold value is generally unknown. In the study, Chan’s (1993) minimal value of the square term of the residual error is taken as the threshold value. Enders and Siklos (2001) suggested applying Akaike information criterion (AIC) or Bayesian information criterion (SBC) to determine the suitable model. As is shown in the result in Table 2, Chan’s M-TAR model is the optimal method for the paired spread of these three bonds. The threshold values of the interest rate of the spread between 20-year bond and 15-year bond, 20-year bond and 10-year bond, 15-year bond and 10-year bond are -0.006, 0.0105 and 0.006 respectively. Regardless of TAR or M-TAR, the null hypothesis ( 1 2 0 ) is refuted and values all pass at least 5% of significance level, namely, no unit-root phenomenon (stationary sequence) exists in the spread. The traditional linear Augmented Dickey-Fuller test result shows the opposite situation. Its result fails to defy null hypothesis, which means that the spread is in non-stationary sequence. In terms of model selection, ADF unit-root test value is more than that of TAR or M-TAR, which means either TAR or M-TAR is superior to the traditional ADF. The spread adjustment results from the threshold. In addition, it can be clearly discovered in Figure 1 that the movement of the interest rate during these three terms is of high correlation, namely, cointegration. The spread shall be a stationary sequence. It is held in the study that the unit-root test in traditional linear hypothesis cannot catch the dynamic and asymmetrical spread adjustment. The test method, if hastily applied, - 13 - may result in deviation. When the result and the asymmetrical effect of the sequence proposed by Pippenger and Goering (1993), Balke and Fomby (1997), Enders and Granger (1998), Enders and Siklos (2001) exist, the traditional unit-root test method and cointegration test will lead to the low power of the test result and the consistency in deviation. <Insert Table 2> When the spread is confirmed to be in stationary state, spread adjustment is conducted to check the asymmetry. It is shown in the F test ( 1 2 ) in 7th line of Table 2 that both TAR and M-TAR test methods refuse null hypothesis and pass the signification level of at least 5%. Therefore, in Table 2 we find the spread is in stationary sequence and the adjustment is conducted in asymmetrical manner. 5.2 Cointegration test on asymmetry Table 3 provides the result of cointegration test on term structure of interest rate. The first and second lines are respectively TAR and M-TAR model developed by Enders and Siklos (2001). The third line is the cointegration test by Engle-Granger(E-G). Its null hypothesis is to check the cointegration between yield rates of government bonds at diverse maturities within the framework of expectation theory. It is found in Table 3 that M-TAR-C model with AIC and SBC as selection standards is superior to the M-TAR and TAR model with the threshold value of zero. Three kinds of nonlinear asymmetrical models, without any exception, refute null hypothesis ( 1 2 0 ) and allow the significance level of at least 5%, that is to say, the cointegration exists in yields of the long-term government bonds at diverse maturities. The spreads between 20-year bond and 15-year bond, 20-year bond and 10-year bond and 15-year bond and 10-year bond are respectively -0.01144, 0.00528 and -0.00959. Chan’s (1993) method can be adopted to find out the most suitable threshold value of M-TAR, namely, the sum of squares of residual error of the model suitable for the threshold value is the minimal. Null hypothesis cannot be refuted when the traditional linear E-G cointegration test is conducted, which means it cannot support the expectation theory of term structure of interest rate. It can be obviously discovered in Figure 1 that there shall be obvious cointegration between the yield rates of the government bonds with different durations. The cointegration test of the traditional linear hypothesis cannot catch the relation between the bonds with long terms. In the study Johansen (1988, 1991) cointegration test was conducted and the - 14 - test result was the same as that of the E-G cointegration test. Both tests cannot refuse null hypothesis, namely, there is no cointegration between government bonds. The conclusion is consistent with the result of the aforementioned unit-root test. The unit-root test in the traditional linear method indicated the spread is in non-stationary sequence while the result of the unit-root test through M-TAR and TAR models selected on the basis of judgment standards: AIC and SBC shows that the spread is in stationary sequence. Therefore, nonlinear dynamic adjustment model is obviously superior to the traditional linear model. <Insert Table 3> In confirming whether the yield rates of government bonds with different durations are subject to nonlinear adjustment, the question of whether the adjustment is asymmetrical or not can be resolved by referring to Table 3. TAR model cannot refuse null hypothesis 1 2 while M-TAR and M-TAR-C refuse null hypothesis and satisfy the significance level of at least 5%. In terms of judgment standards AIC and SBC, M-TAR-C model is best, next is M-TAR model and the last is TAR model. In addition, it can be further discovered in M-TAR and M-TAR-C model that in the asymmetrical isostatic adjustment, the adjustment in the negative direction moves faster than that in the positive direction. According to the actual data, Figure 1 shows the interest rate began to fall since 2000 and it fell very sharply. M-TAR model can catch the feature of the sharp data adjustment. It is observed from Figure 2 that the spread between the government bonds with different durations showed comparatively stable state at the early period in 2000. However, as the central bank continued to lower all kinds of discount rates, interest rate started to fluctuate more violently, which led to the wide fluctuation in the spread. At the second half of 2003, the global economy began to pick up. Although the central bank still adopted the flexible monetary policy, the market as a whole expected the central bank to raise interest rate. As interest rate and the yield rate of government rate rebounded, the fluctuation in interest rate diminished. Therefore M-TAR model can be utilized in the period of wide fluctuation in interest rate to catch the dynamic adjustment of term structure of interest rate. In confirming the presence of the threshold cointegration asymmetry in the yield rate of long-term government bonds, namely, nonlinear relationship4, exists 4 The correlations between short-term, medium-term and long-term interest rates are also considered in the empirical research in the current study. The research targets are interbank call loan rate, 5-year and 10-year government bonds. The results also support expectation theory. - 15 - in the long-term equilibrium between interest rates, the next section will address Threshold Error-Correction Model for the cointegration between interest rates. 5.3 Threshold Error-Correction Model When threshold asymmetrical adjustment cointegration is found between interest rates, we can apply Threshold Error-Correction Model (TECM) to predict the short-term dynamic and long-term balanced relations in the term structure of interest rate. The error-correction model for threshold asymmetry is shown as follows (Enders and Granger, 1998; Enders and Siklos, 2001): Rit 1 Z t1 2 Z t1 ki i 1 Where, i R1t i k2 R i 2t i t (11) i 1 Rit refers to bond yield, Z t1 I t ˆ t 1 , Z t1 (1 I t )ˆ t 1 . When t 1 threshold value, I t 1. On the contrary, when t 1 threshold value, I t 0 and v t the interference item of white noise. The model is selected in accordance with the judgment standards AIC and SBC. From Table 3 we find Chan’s (1993) M-TAR-C model is optimal. The current study adopted M-TAR-C error-correction model to estimate the short-term and long-term balanced relationship between the bonds with different maturities and Granger-Causality was used to test R1,t i or R2,t i . It can be seen from Table 4 that the adjustment to long-term balanced interest rate shows asymmetry (namely, null hypothesis of H 0 : 1 2 is refused) when the difference between the yield rates of these three long-term government bonds satisfies the significance level of 1%. On the other hand, in the relationship between short-term yield rates, we found that mutual influence happens between 20-year and 10-year bonds, as well as between 15-year and 10-year bonds, that is to say, the null hypotheses of H 0 : 1 2 3 4 0 and H 0 : 1 2 3 4 0 are refused. But null hypothesis of H 0 : 1 2 0 and H 0 : 1 2 0 cannot be refused in terms of the relationship between 20-year and 15-year bonds, which represents these two bonds won’t exert mutual influence. The reason can be easily found. Although these three bonds are mainly traded in Taiwan bond market, the hottest is 10-year bond. The 10-year government bond is characterized by the shorter duration and traders suffer less risk5, when the same unit interest rate fluctuates. These factors determine that the Price fluctuation in bond is expressed in terms of duration: P P * R * MD , where P refers to bond price, △R the fluctuation in interest rate and MD Modified Duration which is duration/1+R. It is hypothsized that when the prices of the bonds with different maturities change in the same way as the 5 - 16 - 10-year bond is of high negotiability in the market and naturally becomes the leading indicator of the fluctuation in the yield rate of government bond market. Since the bonds with other maturities will vary with the adjustment to 10-year government bond, the 10-year government bond bears a close relation to the short-term balanced adjustment to 15-year government bond and 20-year government bond. In the long-term balanced relationship, we find that in the process of interest rate adjustment, either the upward or downward adjustment to the threshold value of 10-year, 15-year and 20-year government bond refuse null hypothesis and satisfy the significance level of 1%. 10-year, 15-year and 20-year government bonds are closely related in terms of long-term balanced adjustment. With regard to the long-term balanced adjustment of 20-year and 15-year bonds, these bonds won’t show feedback effect unless the interest rate fluctuates below the threshold value ( =-0.01144). Since the interest rate sharply slows down during the sampling period, when the interest rate of 15-year bond fluctuates more violently than that of 20-year bond, the price conduct direction is from the shorter 15-year bond to the longer 20-year bond. When it is above threshold value, as a result of the poor global economic environment during the sampling period, the short-term fluctuation in interest rate didn’t mean that the interest rate would go up after hitting the bottom. Probably it was merely a rebounce. Therefore, the fluctuation in interest rate of the two bonds above the threshold value failed to generate price feedback effect. After combining the balance relationships between the aforementioned interest rates of government bonds, 10-year bond respectively has price transmission effect over 15-year and 20-year government bonds while 20-year bond and 15-year bond won’t exert price transmission effect in the short run. Only under the circumstances of long-term equilibrium and the interest rate fluctuation is lower than threshold value, can the transmission effect be produced. <Insert Table 4> In addition, it can be seen from Table 4 that in accordance with the factors for the adjustment between yield rates of government bonds: 1 (Z t1 ) and 2 (Z t1 ) , the adjustment between interest rates appears asymmetrical. Between 15-year and 20-year government bonds, the downward adjustment to the interest rate is greater than the upward adjustment, that is to say, the decrease in yield rate will be more than the interest rate changes, the change in bond price is in relation to the duration. The longer bond duration, the greater the change in the profit and loss as a result of the fluctuation in the unit interest rate. - 17 - increase under the long-term equilibrium. From Table 4 we can discover that when the interest rate goes down, the factors for the adjustment to 15-year bond and 20-year bond will generate significant effect while the factor won’t generate significant effect as the interest rate rises. Furthermore, it is discovered that 15-year bond and 20-year bond, when at the threshold value, will exert significant effect over the 10-year bond. When below the threshold value, the 10-year bond will produce significant influence over the 15-year and 20-year bond. The reason is out there that when the interest rate goes down, the government bond with 10-year duration will respond first, which will exert price transmission effect to stimulate the longer 15-year and 20-year government bonds. On the contrary, when the expected interest rate is on the increase, 15-year and 20-year government bond will be first sold in the market in order to reduce the loss of the government bonds with longer duration. Then 10-year bond with shorter duration will be sold. Therefore the adjustment to yield rate of the government bond is asymmetrical. As far as investors and traders are concerned, they can follow M-TAR model to use the relative change in the yield rate of the bond with different durations to adjust the content of their investment portfolio. The relative change in long-term and short-term yield rates will influence the shape of yield curve. If the increase in yield rate of 20-year bond is expected to surpass the threshold value (0.00528) of 10-year bond (namely, the gradient of the yield curve grows sharper), then the proportion of the 10-year bond in the investment portfolio shall be enhanced while that of 20-year bond be lowered6. In the same principle, if the increase in yield rate of 15-year bond is 6 The proportion of long-term bonds can be adjusted by using duration matching analysis as follows: Suppose the investment portfolio consists of two kinds of government bonds with different maturities, x refers to the proportion of the government bond with the first maturity. D1 and D2 refer to the durations for these two bonds. D p ---the duration of investment portfolio is: xD1 (1 x) D2 D p , then the holding proportion x ( D2 D p ) /( D2 D1 ) . If the interest rate sensitivity of investment portfolio is measured on the basis of duration, when threshold value is exceeded, the duration of investment portfolio shall be based on to adjust the proportion of each bond in the investment portfolio. In addition, if the risk of the bond investment portfolio is considered, then variance matching is applied to determine the holding proportion. Suppose two bonds (measured by undulatory property) . 1 and 2 respectively represent the risks of refers to the coefficient of relation between two kinds of bonds, then the variance for investment portfolio is: V ( RP ) x 212 (1 x)2 22 2 x(1 x) 1 2 . When threshold value is modified and the risk is considered, the most appropriate ratio of the bond investment portfolio. Primary differentiation of V ( RP ) and holding proportion x is zero. The most appropriate ratio of government bonds can be - 18 - expected to be below the threshold value (-0.00959) of 10-year bond (namely, the gradient of the yield curve grows slower), the proportion of the 10-year bond in the investment portfolio shall be enhanced while that of 15-year bond be lowered. The same principle applies to the investment portfolio of 15-year bond and 20-year bond. It is important to effectively adjust important portfolio. As regards fund mangers, they can not only work out investment decision based on expectation on interest rate, but also apply M-TAR model to analyze whether the spreads of all kinds of bonds deviate from the equilibrium. The analysis result constitutes the basis for investment. Under the circumstances in which the credit standings of government bonds are the same, there exists certain spread between the government bonds with different maturities. When the spread deviates from the balanced level, then arbitrage can be done by purchasing the underestimated bonds and selling out the overestimated bonds. The profits can be obtained when the spread is restored to the equilibrium. Particularly, in time of arbitrage on the bonds with diverse maturities, if bond sell-out is restricted in the market, the asymmetry will result from raising and lowering the bond price. Additionally, in case of the transaction cost, the price adjustment will generate asymmetrical effect. As far as financial institutions are concerned, the understanding of the adjustment to term structure of interest rate can contribute to the correct evaluation on the values of risk-weighted financial assets and interest rate derivatives, thus reducing the financial losses possibly incurred by the fluctuation in interest rate. Since the banks ’long-term loans with the fixed interest rates originate from the short-term savings, the yield curve risk will rise when yield curve suffers unexpected change and exerts adverse influence over the banks’ interest return and economic value. The good command of the variation in the term structure of interest rate will help banks reduce the risk of timing difference between the maturity (in terms of the fixed interest rate) and re-pricing (in terms of floating interest rate) in order to reduce the return loss as a result of the fluctuation in interest rate. Besides, in terms of financial institutions holding large quantities of government bonds, if they can sense figured out as follows: x 22 1 2 12 22 2 1 2 When the variation in yield rate exceeds the threshold value, the aforementioned two methods are adopted to achieve the most appropriate investment portfolio. - 19 - the dynamic change in interest rate, they can change the traditional conservative strategy (from the purchase date to the maturity) of government bond investment. If the buy-and-hold strategy is adopted, capital gains and interest of the investment portfolio as well as the reinvested income of the principal will be subject to the fluctuation in interest rate. When the interest rate fluctuates violently, if no hedging measures are taken, the value of the investment portfolio in which government bonds are included will undergo the sharp increase or decrease. Therefore, M-TAR model can be adopted to sense the dynamic adjustment to interest rate and enhance the performance of the hedging and evaluation of the investment portfolio. 6. Conclusions A number of empirical studies have revealed that financial sequence data is often subject to nonlinear dynamic adjustment. Enders and Granger (1998) discovered that when the asymmetrical adjustment happens to economic variables, the traditional linear unit-root test and cointegration test lead to the low power. The previous studies have confirmed that in the term structure of interest rate, the premium will vary with time. The recent researches indicated that under the long-term equilibrium, when the premium deviates from the equilibrium, the positive and negative dynamic adjustment will produce asymmetry (Chan and Cheung, 2005). The traditional hypothesis of linear symmetrical adjustment will lead to the deviation in estimation (Enders and Granger, 1998; Enders and Siklos, 2001; van Dijk and Franses, 2000; Tzavalis, 1999; Andreou et al., 2000; Coakley and Fuertes 2002). The present study reviewed the adjustment to the term structure of interest rate of long-term government bonds in Taiwan government bond market from 2000 to the end of 2003. During the research period, Taiwan chanced to suffer the deteriorating economy as a result of both external and internal impacts. For the purpose of stimulating economy, central bank kept lowering all kinds of discount rates for 13 times, which led the interest rate to reach the historical new-low. If the linear symmetry is hypothesized in the adjustment process, error may be generated in model prediction. In the present study we adopted the asymmetrical TAR and M-TAR unit-root and cointegration tests developed by Enders and Granger (1998), Enders and Siklos (2001) and used nonlinear method to test whether term structure of interest rate is consistent with the expectation hypothesis. It departs from the traditional expectation hypothesis in which the premium between different interest rates remains fixed. The empirical - 20 - outcome revealed Taiwan’s term structure of long-term interest rate is consistent with the expectation theory that risk premium is subject to dynamic and asymmetric adjustment. In the study we analyzed 10-year, 15-year and 20-year government bonds in Taiwan and discovered the phenomenon that when interest rate goes up or down, 10-year government bond exerts price transmission effect to 15-year and 20-year government bonds and the spread is subject to asymmetrical dynamic adjustment in both the long and short run. On the other hand, 15-year and 20-year government bonds still have price transmission effect as the interest rate drops. But the transmission effect appears insignificant when the interest rate rises. Therefore, the asymmetrical error-correction model can be applied to capture the dynamic adjustment to interest rate. - 21 - References Lin, C.C., Hong, M.W., Guan, J.M., 1998, “The dynamic behavior of short-term interest rate in Taiwan: the application of regime-switching model”, Taiwan Economic Review, vol. 30: pp.29-55. Shen, J.H.(1993), “The Term Structure of Taiwan Money Market Rates and Rational Expections”, Paper delivered at 1993 Chinese Finance Association Annual Meetings(ROC): pp.227-307. Chuang, W.R., Duan, H.W.(1996), “An Empirical Study on Expectation Model of Taiwan Money Market and Efficient Market Hypothesis: critical test on variance”, Tamkang Periodical, issue 35: pp.207-221. Andreou, E., Osborn, D. R. and Sensier, M., 2000, “A Comparison of the Statistical Properties of Financial Variables in the USA, UK and Germany over the Business Cycle”, The Manchester School, 68, 396-418. Balke, N.S. and Fomby, T., 1997, “Threshold Cointegration”, International Economic Review, 38, 624-643. Campbell, J.Y. and Shiller, R.J., 1987, “Cointegration and Test of Present Value Models”, Journal of Political Economy, 95, 1062-1088. Caner, M. and Hansen, B., 2001, “Threshold autoregression with a unit-root”, Econometrica, 69, 1555-1596. Chan, K.S., 1993, “Consistency and Limiting Distribution of the Least Squares Estimator of a Threshold Autoregressive Model”, The Annals os Statistics, 21, 520-533. Chan, W. S. and Cheung, S. H., 2005, “A bivariate threshold time series model for analyzing Australian interest rates”, Mathematics and Computers in Simulation, 68, 429-437. Coakley, J. and Fuertes, A. M., 2001, “Nonlinearities in Excess Foreign Exchange Returns”, The Manchester School, 69, 623-642. Coakley, J. and Fuertes, A. M., 2002, “Asymmetric Dynamics in UK Real Interest Rates”, Applied Financial Economics, 12, 379-387. Cook. T. and Hahn, T., 1990, “Interest rate expectations and the slope of the money market yield curve”, Fed. Reserve Bank Richmond Econ. Rev. 76, 3-26. Delong, J. B. and Summers, L.H., 1986, “Are Business Cycles Symmetrical”, The American Business Cycle, Chicago, IL, University of Chicago Press, 166-178. Dickey, D.A. and Fuller, W.A., 1979, “Distribution of the Estimators for Autoregression Time Series with a Unit Root”, Journal of American Statistical Association, 74, 427-432. Enders, W. and Dibooglu, S., 2002, “Long-run purchasing power parity with - 22 - asymmetric adjustment”, Southern Economics Journal 26, 101-110. Enders, W. and Granger, C.W.F., 1998, “Unit-Root Tests and Asymmetric Adjustment With an Example Using the Term Structure of Interest Rates”, Journal of Business Economics & Statistics, 19, 166-176. Enders, W. and Siklos, P.L., 2001, “Cointegration and Error Correction: Representation, Estimation, and Testing”, Econometrica, 55, 251-276. Engle, R. F. and Granger, C. W. J., 1987, “Cointegration and Error Correction: Representation, Estimation, and Testing”, Econometrica, 55, 251-276. Falk, B., 1986, “Further Evidence on the Asymmetric Behaviour of Economic Time Series over the Business Cycle”, Journal of Political Economy, 94, 1096-1109. Fisher, I., 1930, The Theory of Interest, Macmillan, New York. Gerlach, S. and Smets, F., 1997, “The Term Structure of Euro-rates: Some Evidence in Support of the EH”, Journal of International Money and Finance, 16(April), 305-321. Hall, A. D., Anderson, H. M., and Granger, C. W. J., 1992, “A cointegration analysis of treasury bill yields”, Review of Economics and Statistics 74, 116-126. Hardouvelis, G.A., 1988, “The Predictive Power of The Term Structure During Recent Monetary Regimes”, Journal of Finance 43, 339-356. Iyer, S., 1997, “Time-varying term premia and the behavior of forwardd interest rate prediction errors”, The Journal of Financial Research, 20, 504-507. Johansen, S., 1988, “Statistical Analysis of Cointegration Vectors”, Journal of Economic Dynamics and Control, 12, 231-254. Johansen, S., 1991, “Estimation and Hypotheses in Testing of Cointegration Vector in Gaussian Vector Autoregressive Model”, Econometrica, 59, 1551-1580. Johansen, S., 1996, Likelihood-based Inference in Cointegrated Vector Autoregressive Models, Oxford, Oxford University Press. Kuo, S.H. and Enders, W., 2004, “The term structure of Japanse interest rates: The equilibrium spread with asymmetric dynamics”, Journal of The Japanese and International Economies, 18, 84-98. Kwiatkowski, Denis, Phillips, Peter, Schmidt, Peter and Shin, Yongcheol, 1992, “Testing the Null Hypothesis of Stationary Against the Alternative of a Unit Root”, Journal of Econometrics, 54, 159-178. Lutz, F. A., 1940, “The Structure of Interest Rates,” Quarterly Journal of Economics, 55, 36-63. Mankiw, N. G.. and Miron, J.A., 1986, “The Changing Behavior of The Term Structure of Interest Rates”, The Quarterly Journal of Economics 101, 211-228. Neftci, S. N., 1984, “Are Economic Time Series Asymmetric over the Business Cycle”, Journal of Political Economy, 92, 370-328. - 23 - Obstfeld, M. and Taylor, M., 1997, “Nonlinear Aspects of Goods-market Arbitrage and Adjustment: Heckscher’s Commodity Points Revisited”, Journal of Applied Probability, 21, 270-286. Phillips, P.C.B. and Perron, P., 1988, “Testing for a Unit Root in Time Series Regression”, Biometrika, 75, 335-346. Pippenger, M.K. and Goering, G. E., 1993, “A note on the empirical power of unit root tests under threshold processes”, Oxford Bulletin of Economics and Statistics, 55, 473-481. Rudebusch, G..D., 1995, “Federal Reserve Interest Rate Targeting, Rational Expectations, and the Term Structure”, J. Monet. Econ. 35, 245-274. Said, S. E. and Dickey, D. A., 1980, “Testing for unit roots in ARMA models of unknown order”, Biometrika, 71, 599-607. Shiller, R. J., 1979, “The volatility of long-term interest rates and expectations models of the term structure”, Journal of Political Economy, 87, 1190-1219. Sichel, D., 1989, “are Business Cycles Asymmetric? A Correction”, Journal of Political Economy, 97, 1255-1260. Sichel, D., 1993, “Business Cycles Asymmetric: A Deeper Look”, Economic Inquiry, 31, 224-236. Stlvester, C. W. Eijffinger and Haan, Jakob De., 1996, The Political Economy of Central Bank Independence, Special Papers in International Economics, Princeton, Princeton University Press. Terasvirta, T. and Anderson, H. M., 1992, “Characterising Nonlinearities in Business Cycles Using Smooth Transition Autoregressive Models”, Journal of Applied Econometrics, 7, 119-136. Tong, H., 1983, Threshold Models in Non-Linear Time Series Analysis, New York: Springer-Verlag. Tzavalis, E., 1999, “A Common Shift in Real Interest Rates across Countries”, Applied Financial Economics, 9, 365-369. van Dijk, D. and Franses, P. H., 2000, “Non-linear Error-correction Models for Interest Rates in the Netherlands”, Nonlinear Econometric Modeling-Proceedings of the 6th (EC) 2 Meeting, Cambridge, Cambridge University Press, 203-227. - 24 - Figure 1 Government bond yield in 2000 to 2003 7 6 5 4 3 2 1 2000/01/04 2000/11/29 R10 2001/11/30 R15 2002/12/04 2003/12/31 R20 Figure 2 Differential between the 10, 15 and 20-year government bond yield 1.2 1.0 0.8 0.6 0.4 0.2 0.0 -0.2 -0.4 2000/01/04 2000/11/29 R15R10 2001/11/30 R20R10 - 25 - 2002/12/04 2003/12/31 R20R15 Table1 Summary Statistics of Government Bond Yield 10-Year 15-Year 20-Year Mean 3.8602% 4.0965% 4.2673% Max. 6.1020% 6.2000% 6.2500% Min 1.3120% 1.4725% 1.9725% Std. 1.3652% 1.2605% 1.1366% Skewness 0.0042 -0.1267 0.0977 Kurtosis 1.9177 2.0818 1.9448 49.5385*** 38.3699*** 48.699*** 1015 1015 1015 ADF(n) -1.4539(5) -1.4339(7) -1.3694(13) PP -1.4298(13) -1.3855(13) -1.3175(13) 3.5374(25)*** 3.5004(25) *** 3.6067(25) *** 4027.6*** 4025.4*** 4018.2*** J-B Sample Size KPSS L-B(Q=4) Note: 1. J-B denotes the Jerque-Bera Test for Normality 2. L-B(Q=k) denotes the Ljung-Box Test for Autoregressive for residual in k lags. 3. ADF denotes the Augmented Dickey-Fuller unit root test. It depends on AIC(Akaike information criterion) criteria for lags. 4. PP denotes Phillips and Perron unit root test.,KPSS denotes Kwiatkowskiet et al. unit root test. 5. PP and KPSS based on Newey and West Bandwidth criteria for lags 6. The ***, **, and * indicate significance at the 0.01, 0.05 and 0.1 levels, respectively. - 26 - Table 2 Test for a unit root in the yield spread(sample period: 2000/1/4 to 2003/12/31,n=1015) Model ρ1 ρ2 AIC/SBC1 Φ2 ρ1=ρ23 Q(4)4 -0.0062 (-1.7323)5 0.0016 (0.3060) -0.0667 (-1.8739) 0.0055 (1.1536) -0.2532*** (-3.6594)6 -0.0162*** (-2.8254) -145.0008/ -120.4073 -136.8217/ -112.2282 -135.5390/ -115.8570 -147.3721/ -122.7787 8.1169*** 12.7261*** 3.9986** 4.5177** 9.3171*** 15.1184*** 3.8541 (0.4261) 2.9324 (0.5692) 2.8285 (0.5870) 3.3462 (0.5017) 371.5835/ 386.3455 355.9690/ 370.7311 375.0354/ 394.7181 351.2924/ 366.0544 46.6405*** 91.2698*** 9.5237*** 17.8048*** 11.9046*** 22.5608*** 138.9869/ 163.5803 139.0443/ 163.6377 142.4780/ 162.1607 135.6527/ 160.2462 16.4725*** 30.5516*** 11.3305*** 20.2912*** 13.0588*** 23.7398*** Lag (20-year and15-year yield rate) TAR 3 M-TAR 3 ADF 3 M-TAR-C7 ( =-0.006) 3 -0.0284*** (-4.2852) (20-year and 10-year yield rate) TAR 1 M-TAR 1 ADF 1 M-TAR-C ( =0.0105) 1 -0.0028 (-1.2460) 0.0085*** (2.4007) -0.0028 (-1.1837) 0.0147*** (3.3929) -12.0404*** (-9.5911) -0.9600*** (-4.0536) -0.0130*** (-4.0561) 2.9456 (0.5670) 2.7114 (0.6072) 3.6181 (0.4600) 3.4003 (0.4932) (15-year and 10-year yield rate) TAR 3 M-TAR 3 ADF 3 M-TAR-C ( =0.006) 3 -0.0052 (-1.4688) -0.0131*** (-2.3976) -0.006 (-1.6837) 0.0202*** (3.1658) -0.6787*** (-5.5703) -0.0261*** (-4.4961) -0.0227*** (-4.5205) 1.3552 (0.8519) 2.7460 (0.6012) 3.1941 (0.5260) 3.4656 (0.4813) Note: 1. 7. AIC=T * ln(RSS+2 * n;SBC=T * ln(RSS)+n * ln(T),where n = number of regressors and T = number of usable observations. RSS is residual sum of squares. Entries in this column are sample F-statistics for testing the null of 1 2 0 . This test follows a non-standard distribution so the test statistics are compared with critical values reported by Enders and Siklos(2001) Entries in this column are sample F-statistics for the null hypothesis that adjustments are symmetric. The critical values are reported by Enders and Siklos(2001). Q(4) is the Ljung-Box Q-statistic for the joint hypotheses of no serial correlation among the first residuals. Entries in the brackets of this column are the t-statistic for the null hypothesis 1 0 . Critical values are taken from Enders and Granger(1998). Entries in the brackets of this column are the t-statistic for the null hypothesis 2 0 . Critical values are taken from Enders and Granger(1998). The C denotes the model estimation using a consistent threshold estimate. 8. The ***, **, and * indicate significance at the 0.01, 0.05 and 0.1 levels, respectively. 2. 3. 4. 5. 6. - 27 - Table 3 Test for cointegration of the term structure (sample period: 2000/1/4 to 2003/12/31,n=1015) Model ρ1 ρ2 AIC/SBC1 Φ2 ρ1=ρ23 Q(4)4 -0.0221*** (-2.5370)5 -0.0077 (-0.8811) -0.0221 (-3.5233) -0.0048 (-0.6598) -0.0172** (-1.9172)6 -0.0320*** (-3.6110) -266.4757/ -241.8822 -270.2111/ -245.6176 -272.8171/ -253.1305 -281.4876/ -256.8941 4.9792** 0.1565 6.8596*** 3.8809** 12.5783*** 15.2080*** 2.6222 (0.6229) 2.6618 (0.6159) 2.3339 (0.675) 2.7071 (0.6080) 4.3976** 0.2879 8.1041*** 7.6390*** 10.3161*** 12.0261*** 5.3306*** 0.7712 10.6103*** 11.2280*** 12.7723*** 15.5196*** Lag (20-year and15-year yield rate) TAR 3 M-TAR 3 E-G 3 M-TAR-C7 ( =-0.01144) 3 -0.0584*** (-4.9800) (20-year and 10-year yield rate) TAR 0 M-TAR 0 E-G 0 M-TAR-C ( =0.00528) 0 -0.0226** (-2.3357) -0.0014 (-0.1581) -0.0232 (-3.6313) 0.0073 (0.7410) -0.0157* (-1.8275) -0.0367*** (-4.0228) -0.0367*** (-4.4814) 102.1033/ 111.9466 94.7663/ 104.6096 95.7829/ 100.7046 90.4128/ 100.2561 2.7339 (0.6033) 1.8473 (0.7638) 2.8808 (0.5780) 1.7387 (0.7837) (15-year and 10-year yield rate) TAR 3 M-TAR 3 E-G 3 M-TAR-C ( =-0.00959) 3 -0.0176* (-1.7840) 0.0012 (0.1172) -0.0236 (-3.2137) -0.0029 (-0.3220) -0.0304*** (-2.7683) -0.0475*** (-4.6034) -0.0631*** (-5.0490) 38.6585/ 63.2520 28.2121/ 52.8056 23.4666/ 49.1532 23.9556/ 48.5491 2.4107 (0.6607) 1.8925 (0.7555) 2.7806 (0.5950) 2.2731 (0.6857) Note: 1. 2. 3. 4. 5. 6. 7. 8. AIC=T * ln(RSS+2 * n;SBC=T * ln(RSS)+n * ln(T),where n = number of regressors and T = number of usable observations. RSS is residual sum of squares. Entries in this column are sample F-statistics for testing the null of 1 2 0 . This test follows a non-standard distribution so the test statistics are compared with critical values are taken from by Enders and Siklos(2001) Entries in this column are sample F-statistics for the null hypothesis that adjustments are symmetric. The critical values are reported by Enders and Siklos(2001). Q(4) is the Ljung-Box Q-statistic for the joint hypotheses of no serial correlation among the first residuals. Entries in the brackets of this column are the t-statistic for the null hypothesis 1 0 . Critical values are taken from Enders and Granger(1998). Entries in the brackets of this column are the t-statistic for the null hypothesis 2 0 . Critical values are taken from Enders and Granger(1998). The C denotes the model estimation using a consistent threshold estimate. The ***, **, and * indicate significance at the 0.01, 0.05 and 0.1 levels, respectively. - 28 - Table 4 The estimated asymmetric error-correction models for the term spread △R10 △R20 △R10 △R15 C -0.0030(-1.7620)* -0.0031(-2.1809)** -0.0026(-1.4983) -0.0029(-1.8795)* △R10(-1) -0.0782(-1.6971)* -0.0403(-1.0624) -0.0913(-1.7551)* -0.0168(-0.3625) △R10(-2) 0.1478(3.1915)*** 0.1012(2.6523)*** 0.1457(2.7989)*** 0.1376(2.9558)*** △R10(-3) -0.1099(-2.3645)** -0.1179(-3.0820)*** -0.1321(-2.5368)** -0.0344(-0.7347) △R10(-4) 0.1434(3.008)*** 0.1081(2.8111)*** 0.0700(1.3346) 0.1235(2.6281)*** △R20(-1) 0.1537(2.7452)*** 0.1360(29510)*** 0.1371(2.3748)** 0.0648(1.2558) △R20(-2) -0.2004(-3.5709)*** -0.1150(-2.4840)** -0.1545(-2.6569)*** -0.1433(-2.7627)*** △R20(-3) 0.0745(1.3221) 0.0609(1.3126) 0.0957(1.6533)* -0.0027(-0.0512) △R20(-4) 0.1153(-1.0418)** -0.0839(-1.8061)* -0.0096(-0.1592) -0.0655(-1.2569) 1 -0.0264(-1.9147)* 0.0081(0.5961) -0.0318(-2.2592)** -0.0219(-1.5736) 2 -0.0036(-0.3064) -0.0207(-1.7962)* 0.0237(1.1497) 0.0142(0.7357) H 0 : 1 2 3 4 0 7.3891*** H 0 : 1 2 3 4 0 6.3303*** 4.5132*** 4.0484*** H 0 : 1 2 3 4 1 0 5.9323*** 4.4151*** H 0 : 1 2 3 4 2 0 6.4791*** 3.7022*** H 0 : 1 2 3 4 1 0 5.9819*** 4.4149*** H 0 : 1 2 3 4 2 0 5.0657*** 3.5065*** R10 H0 : 1 2 R20 H0 : 1 2 R10 H0 : 1 2 R15 H0 : 1 2 19.6229*** 14.5081*** 13.4018*** 8.4452*** Q(4) 0.0309 0.1775 0.0627 0.0504 Note: 1. 2. The ***, **, and * indicate significance at the 0.01, 0.05 and 0.1 levels, respectively. t-statistics are in parentheses. Threshold Error-Correction model: ki kI i 1 i 1 Rit 1 Z t1 2 Z t1 i R1t i i R2t i t Where t 1 3. 4. 5. Rit is term yield, and vt is a Z t1 I t ˆ t 1 , Z t1 (1 I t )ˆ t 1 such that I t 1 white-noise disturbance The threshold is 0.00528 in 10-year and 20-year error-correction model. The threshold is -0.00959 in 10-year and 20-year error-correction model. Lag-length selection is based on AIC (Akaike information criterion). - 29 - if t 1 , I t 0 if Table 4 The estimated asymmetric error-correction models for the term spread (continued) △R15 △R20 C -0.0035(-2.2307)** -0.0035(-2.4438)** △R15(-1) 0.0068(0.1251) 0.0338(0.6890) △R15(-2) 0.0228(0.4242) 0.0220(0.4483) △R20(-1) 0.0597(1.0179) 0.0523(0.9762) △R20(-2) -0.0622(1.0656) -0.0382(-0.7147) 1 0.0039(0.3305) 0.0142(1.1785) 2 -0.0437(-2.3051)** -0.0739(-3.8315)*** H 0 : 1 2 0 0.3217 H 0 : 1 2 0 1.1115 H 0 : 1 2 1 0 0.6136 H 0 : 1 2 2 0 5.2695*** H 0 : 1 2 1 0 0.7852 H 0 : 1 2 2 0 2.6143** R15 H0 : 1 2 R20 H0 : 1 2 4.5300** 15.0503*** Q(4) 2.5441 1.8782 Note: 1. 2. The ***, **, and * indicate significance at the 0.01, 0.05 and 0.1 levels, respectively. t-statistics are in parentheses. Threshold Error-Correction model: ki kI i 1 i 1 Rit 1 Z t1 2 Z t1 i R1t i i R2t i t Where t 1 3. 4. Rit is term yield, and vt is a Z t1 I t ˆ t 1 , Z t1 (1 I t )ˆ t 1 such that I t 1 white-noise disturbance The threshold is -0.01144 in 15-year and 20-year error-correction model. Lag-length selection is based on AIC (Akaike information criterion). - 30 - if t 1 , I t 0 if